mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

ggml-cpu: arm64: q5_K repack gemm and gemv (and generic) implementations (dotprod) (#19356)

Browse Source 
* Generic GEMV and boilerplate for q5_K dotprod
* Generic GEMM and boilerplate for q5_K dotprod
* ARM64 q5_K dotprod GEMM
* ARM64 q5_K dotprod GEMV
This commit is contained in:
Alberto Cabrera Pérez
2026-02-23 12:42:52 +00:00
committed by GitHub
parent 2b6dfe824d
commit bc160d3582
4 changed files with 642 additions and 201 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cpu/repack.cpp
+235 -200
View File
@@ -450,6 +450,208 @@ static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int n,
}
}
template <int M, int N>
static void ggml_gemv_q5_K_NxM_q8_K_generic_impl(int n,
float * GGML_RESTRICT s,
size_t bs,
const void * GGML_RESTRICT vx,
const void * GGML_RESTRICT vy,
int nr,
int nc) {
constexpr int blocklen = M;
constexpr int ncols_interleaved = N;
const int qk = QK_K;
const int nb = n / qk;
static const uint32_t kmask1 = 0x3f3f3f3f;
static const uint32_t kmask2 = 0x0f0f0f0f;
static const uint32_t kmask3 = 0x03030303;
assert(n % qk == 0);
assert(nc % ncols_interleaved == 0);
UNUSED(bs);
UNUSED(nr);
float sumf[ncols_interleaved];
float sum_minf[ncols_interleaved];
uint32_t utmp[32];
int sumi1;
int sumi2;
int sumi;
const block_q8_K * a_ptr = (const block_q8_K *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) {
sumf[j] = 0.0;
sum_minf[j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int sb = 0; sb < 8; sb++) {
memcpy(utmp + sb * 4, b_ptr[l].scales + sb * K_SCALE_SIZE, K_SCALE_SIZE);
utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
utmp[sb * 4 + 2] = uaux_0;
utmp[sb * 4 + 0] &= kmask1;
}
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
constexpr int scale_stride = 32;
uint8_t * scales_0 = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride;
uint8_t * scales_1 = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride + 16;
const int qh_shift = (k / (32 / blocklen)) * 2;
for (int j = 0; j < ncols_interleaved; j++) {
sumi1 = 0;
sumi2 = 0;
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
const int qh_idx = (k * blocklen + i) % 32;
const int qh_chunk = qh_idx / blocklen;
const int qh_pos = qh_idx % blocklen;
const int b_qh_offset = qh_chunk * (blocklen * ncols_interleaved) + j * blocklen + qh_pos;
const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
const uint8_t h0 = (qh_val >> qh_shift) & 1;
const uint8_t h1 = (qh_val >> (qh_shift + 1)) & 1;
const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
const int q8_offset = (k / (32 / blocklen)) * 64 + (k % (32 / blocklen)) * blocklen + i;
sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
sumi2 = (v1 * a_ptr[l].qs[q8_offset + 32]);
sumi1 = sumi1 * scales_0[j];
sumi2 = sumi2 * scales_1[j];
sumi += sumi1 + sumi2;
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
}
}
for (int sb = 0; sb < 8; sb++) {
uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
for (int j = 0; j < ncols_interleaved; j++) {
sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) *
GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
}
}
}
for (int j = 0; j < ncols_interleaved; j++) {
s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
}
}
}
template <int M, int N>
static void ggml_gemm_q5_K_NxM_q8_K_generic_impl(int n,
float * GGML_RESTRICT s,
size_t bs,
const void * GGML_RESTRICT vx,
const void * GGML_RESTRICT vy,
int nr,
int nc) {
constexpr int blocklen = M;
constexpr int ncols_interleaved = N;
const int qk = QK_K;
const int nb = n / qk;
static const uint32_t kmask1 = 0x3f3f3f3f;
static const uint32_t kmask2 = 0x0f0f0f0f;
static const uint32_t kmask3 = 0x03030303;
assert(n % qk == 0);
assert(nr % 4 == 0);
assert(nc % ncols_interleaved == 0);
float sumf[4][ncols_interleaved];
float sum_minf[4][ncols_interleaved];
uint32_t utmp[32];
int sumi1;
int sumi2;
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumf[m][j] = 0.0;
sum_minf[m][j] = 0.0;
}
}
for (int l = 0; l < nb; l++) {
for (int sb = 0; sb < 8; sb++) {
memcpy(utmp + sb * 4, b_ptr[l].scales + sb * K_SCALE_SIZE, K_SCALE_SIZE);
utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
utmp[sb * 4 + 2] = uaux_0;
utmp[sb * 4 + 0] &= kmask1;
}
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
constexpr int scale_stride = 32;
uint8_t * scales_0 = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride;
uint8_t * scales_1 = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride + 16;
const int qh_shift = (k / (32 / blocklen)) * 2;
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi1 = 0;
sumi2 = 0;
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
const int qh_idx = (k * blocklen + i) % 32;
const int qh_chunk = qh_idx / blocklen;
const int qh_pos = qh_idx % blocklen;
const int b_qh_offset =
qh_chunk * (blocklen * ncols_interleaved) + j * blocklen + qh_pos;
const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
const uint8_t h0 = (qh_val >> qh_shift) & 1;
const uint8_t h1 = (qh_val >> (qh_shift + 1)) & 1;
const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
const int q8_offset = (k / (32 / blocklen)) * 256 +
(k % (32 / blocklen)) * 4 * blocklen + m * blocklen + i;
sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
sumi2 = (v1 * a_ptr[l].qs[q8_offset + 128]);
sumi1 = sumi1 * scales_0[j];
sumi2 = sumi2 * scales_1[j];
sumi += sumi1 + sumi2;
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
}
}
}
for (int sb = 0; sb < 8; sb++) {
uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
for (int m = 0; m < 4; m++) {
const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
for (int j = 0; j < ncols_interleaved; j++) {
sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) *
GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
}
}
}
}
}
extern "C" {
void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -803,98 +1005,12 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
}
}
void ggml_gemv_q5_K_8x8_q8_K_generic(int n,
float * GGML_RESTRICT s,
size_t bs,
const void * GGML_RESTRICT vx,
const void * GGML_RESTRICT vy,
int nr,
int nc) {
const int qk = QK_K;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
static const uint32_t kmask1 = 0x3f3f3f3f;
static const uint32_t kmask2 = 0x0f0f0f0f;
static const uint32_t kmask3 = 0x03030303;
void ggml_gemv_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
ggml_gemv_q5_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
}
assert(n % qk == 0);
assert(nc % ncols_interleaved == 0);
UNUSED(bs);
UNUSED(nr);
float sumf[8];
float sum_minf[8];
uint32_t utmp[32];
int sumi1;
int sumi2;
int sumi;
const block_q8_K * a_ptr = (const block_q8_K *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) {
sumf[j] = 0.0;
sum_minf[j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int sb = 0; sb < 8; sb++) {
memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
utmp[sb * 4 + 2] = uaux_0;
utmp[sb * 4 + 0] &= kmask1;
}
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
uint8_t * scales_0 = (uint8_t *) utmp + (k / 4) * 32;
uint8_t * scales_1 = (uint8_t *) utmp + (k / 4) * 32 + 16;
const int qh_shift = (k / 4) * 2;
for (int j = 0; j < ncols_interleaved; j++) {
sumi1 = 0;
sumi2 = 0;
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
const int qh_idx = (k * 8 + i) % 32;
const int qh_chunk = qh_idx / 8;
const int qh_pos = qh_idx % 8;
const int b_qh_offset = qh_chunk * 64 + j * 8 + qh_pos;
const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
const uint8_t h0 = (qh_val >> qh_shift) & 1;
const uint8_t h1 = (qh_val >> (qh_shift + 1)) & 1;
const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
const int q8_offset = (k >> 2) * 64 + (k % 4) * blocklen + i;
sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
sumi2 = (v1 * a_ptr[l].qs[q8_offset + 32]);
sumi1 = sumi1 * scales_0[j];
sumi2 = sumi2 * scales_1[j];
sumi += sumi1 + sumi2;
}
sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
}
}
for (int sb = 0; sb < 8; sb++) {
uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
for (int j = 0; j < ncols_interleaved; j++) {
sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) *
GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
}
}
}
for (int j = 0; j < ncols_interleaved; j++) {
s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
}
}
void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
ggml_gemv_q5_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
}
@@ -1494,107 +1610,12 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
}
}
void ggml_gemm_q5_K_8x8_q8_K_generic(int n,
float * GGML_RESTRICT s,
size_t bs,
const void * GGML_RESTRICT vx,
const void * GGML_RESTRICT vy,
int nr,
int nc) {
const int qk = QK_K;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
void ggml_gemm_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
ggml_gemm_q5_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
}
constexpr uint32_t kmask1 = 0x3f3f3f3f;
constexpr uint32_t kmask2 = 0x0f0f0f0f;
constexpr uint32_t kmask3 = 0x03030303;
assert(n % qk == 0);
assert(nr % 4 == 0);
assert(nc % ncols_interleaved == 0);
float sumf[4][8];
float sum_minf[4][8];
uint32_t utmp[32];
int sumi1;
int sumi2;
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumf[m][j] = 0.0;
sum_minf[m][j] = 0.0;
}
}
for (int l = 0; l < nb; l++) {
for (int sb = 0; sb < 8; sb++) {
memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
utmp[sb * 4 + 2] = uaux_0;
utmp[sb * 4 + 0] &= kmask1;
}
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
uint8_t * scales_0 = (uint8_t *) utmp + (k / 4) * 32;
uint8_t * scales_1 = (uint8_t *) utmp + (k / 4) * 32 + 16;
const int qh_shift = (k / 4) * 2;
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi1 = 0;
sumi2 = 0;
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
const int qh_idx = (k * 8 + i) % 32;
const int qh_chunk = qh_idx / 8;
const int qh_pos = qh_idx % 8;
const int b_qh_offset = qh_chunk * 64 + j * 8 + qh_pos;
const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
const uint8_t h0 = (qh_val >> qh_shift) & 1;
const uint8_t h1 = (qh_val >> (qh_shift + 1)) & 1;
const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
const int q8_offset = (k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i;
sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
sumi2 = (v1 * a_ptr[l].qs[q8_offset + 128]);
sumi1 = sumi1 * scales_0[j];
sumi2 = sumi2 * scales_1[j];
sumi += sumi1 + sumi2;
}
sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
}
}
}
for (int sb = 0; sb < 8; sb++) {
uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
for (int m = 0; m < 4; m++) {
const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
for (int j = 0; j < ncols_interleaved; j++) {
sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) *
GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
}
}
}
}
void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
ggml_gemm_q5_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
}
void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2029,18 +2050,16 @@ static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_in
const int end = QK_K * 4 / blck_size_interleave;
// Interleave Q5_K quants by taking 8 bytes at a time
// Interleave Q5_K quants by taking blck_size_interleave bytes at a time
for (int i = 0; i < end; ++i) {
int src_id = i % 8;
int src_offset = (i / 8) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], blck_size_interleave);
}
// Repeat for low bits 8 bytes at a time as well, since
// Repeat for high bits with the same chunk size, since
// the high bits are interleaved in Q5_K and the index is
// qh_idx = (qs_idx % 32);
// qh_val = qh[qh_idx] >> (qs_idx / 32);
@@ -2049,9 +2068,7 @@ static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_in
int src_offset = (i / 8) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
memcpy(&elems, &in[src_id].qh[src_offset], sizeof(uint64_t));
memcpy(&out.qh[dst_offset], &elems, sizeof(uint64_t));
memcpy(&out.qh[dst_offset], &in[src_id].qh[src_offset], blck_size_interleave);
}
// The below logic is copied over from Q4_K
@@ -2249,7 +2266,7 @@ static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor * t,
const void * GGML_RESTRICT data,
size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_Q5_K);
GGML_ASSERT(interleave_block == 8);
GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
constexpr int nrows_interleaved = 8;
block_q5_Kx8 * dst = (block_q5_Kx8 *) t->data;
@@ -2523,6 +2540,10 @@ template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * da
return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
}
template <> int repack<block_q5_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_q5_K_to_q5_K_8_bl(t, 4, data, data_size);
}
template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
}
@@ -2591,6 +2612,10 @@ template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_q5_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_q5_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
@@ -2654,6 +2679,10 @@ template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_q5_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_q5_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
}
template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
@@ -3068,6 +3097,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
// instance for Q5_K
static const ggml::cpu::repack::tensor_traits<block_q5_K, 4, 8, GGML_TYPE_Q8_K> q5_K_8x4_q8_K;
static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;
// instance for Q6_K
@@ -3130,6 +3160,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
return &q5_K_8x8_q8_K;
}
}
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
if (cur->ne[1] % 8 == 0) {
return &q5_K_8x4_q8_K;
}
}
} else if (cur->type == GGML_TYPE_Q6_K) {
if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
if (cur->ne[1] % 8 == 0) {